ANE Fieldfare

About

ANE Fieldfare is a Swift runtime that runs the Gemma 4 26B-A4B mixture-of-experts model on Apple Silicon Macs. It uses the Neural Engine for prompt prefill, speculative decoding on the GPU, and memory-mapped expert weights so the OS file cache acts as the expert cache. The README reports up to about 97 tokens per second on greedy generation with a 16 GB Mac, compared with 24 to 25 for Turbo Fieldfare, which it builds on. These figures are the author's own measurements.